A generalizable deep learning framework for inferring fine-scale germline mutation rate maps

Abstract Germline mutation rates are essential for genetic and evolutionary analyses. Yet, estimating accurate fine-scale mutation rates across the genome is a great challenge, due to relatively few observed mutations and intricate relationships between predictors and mutation rates. Here we present MuRaL ( Mu tation Ra te L earner), a deep learning framework to predict mutation rates at the nucleotide level using only genomic sequences as input. Harnessing human germline variants for comprehensive assessment, we show that MuRaL achieves better predictive performance than current state-of-the-art methods. Moreover, MuRaL can build models with relatively few training mutations and a moderate number of sequenced individuals, and can leverage transfer learning to further reduce data and time demands. We apply MuRaL to produce genome-wide mutation rate maps for four representative species - Homo sapiens, Macaca mulatta, Arabidopsis thaliana and Drosophila melanogaster , demonstrating its high applicability. As an example, we use improved mutation rate estimates to stratify human genes into distinct groups which are enriched for different functions, and highlight that many developmental genes are subject to high mutational burden. The open-source software and generated mutation rate maps can greatly facilitate related research.